Agentic Search Models with OpenSearch and Elasticsearch
3 hours ago
- #Agentic Retrieval
- #LLM Applications
- #Search Optimization
- Introducing SID-1 model: an agentic LLM designed for search and reranking, offering easy relevance improvement.
- Addresses common search problems: results not at top, noisy results, language mismatch, and vector search limitations.
- SID-1 outperforms models like Gemini 3 Pro, Sonnet 4.5, GPT-5.1 in accuracy, and is 24x faster.
- Agentic retrieval pattern: writes multiple query variants, executes them, picks best results, and iterates.
- SID typically uses 2-3 turns for retrieval plus a reranking turn, optimizing efficiency.
- Cost-effective compared to GPT-5.1, priced similarly to GPT-4o-mini, with lower token usage.
- Implementation integrates into existing search apps via OpenAI-compatible API with minimal code changes.
- Transparent execution: UI updates show progress, queries are batched using OpenSearch's _msearch for efficiency.
- Source code available for experimentation with a Gutenberg dataset, enabling drop-in agentic search.